Upload
others
View
24
Download
0
Embed Size (px)
Citation preview
LINEAR AND NON-LINEAR FINITE ELEMENT ERRORESTIMATION BASED ON ASSUMED STRAIN FIELDS
F. Gabaldon1,∗and J.M. Goicolea1
1 E.T.S. Ingenieros de Caminos, Canales y Puertos, Universidad Politecnica de Madrid,28040 Madrid, Spain
SUMMARY
In this work we propose an error estimator for finite elements in solid mechanics applications. Theestimator evaluates the quality of the solution for standard Galerkin displacement elements, takinginto account the enrichment of the solution with enhanced assumed strain (EAS) mixed elements. Thecontribution of the enhanced strain modes is measured with an energy norm, which is the basis of theestimator.
The error estimator proposed has two interesting advantages. Firstly, it results in a local formulationwhich is evaluated element by element. Secondly, it is easily extended to nonlinear problems. In thiswork, the formulation is developed for linear elasticity, for finite strain elasticity, and von Mises smallstrain plasticity.
Finally, some representative numerical simulations are presented which show the performance ofthe method.
Copyright c© 2000 John Wiley & Sons, Ltd.
key words: Mixed Elements, Error Estimation, Hiperelasticity, Plasticity, Assumed Strain
1. INTRODUCTION
Finite element methods have become one of the main computational tools for applied mechanicsand structural engineering. Many successful applications exist nowadays for both linear andnonlinear problems. This expansion in the field of applications must be accompanied withparallel efforts to assess the quality of the computed results.
The interest of “a posteriori” error estimators lies on their direct applicability to adaptiverefinement techniques. The development of these estimators may be traced back to the worksof Babuska[3]. Since then, several error estimators have been proposed for linear analyses,whose efficiency has been proved in a wide variety of problems.
The development of nonlinear finite element methods is more recent, and a number ofresearch lines remain open. Error estimation applied to this kind of problems has some
∗Correspondence to: F. Gabaldon. Departamento de Mecanica de Medios Continuos y Teorıa de Estructuras.E.T.S. Ingenieros de Caminos, Canales y Puertos. C/ Profesor Aranguren s/n, Ciudad Universitaria, 28040Madrid, Spain. e-mail: [email protected]
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 1
inherent limitations due to the nature of nonlinear behaviour. As examples, one might citenon-unique solutions associated to load path dependency, or bifurcations associated to materialor geometric instabilities such as localisation of deformation or buckling.
In spite of these limitations, some remarkable achievements have been performed in nonlinearfinite element error estimation. Between these are the developments of Ortiz and Quigley [9]in localisation, Johnson [8] in the context of the elastic-plastic model of Hencky, the errorestimator of Barthold et al [4] applied to the elastic-plastic models of Hencky and Prandtl-Reuss, etc. Finally it’s necessary to remark the recent works of Radovitzky and Ortiz [11] inerror estimation for highly non linear problems, including finite deformations in hyperelasticity,viscoplasticity, dynamics, etc.
In what remains of this paper, a new method for error estimations is proposed, applicable tolinear and nonlinear problems. It is based on the improved approximation qualities of enhancedassumed strain (EAS) mixed formulations such as have been proposed in [13, 15]. The errorestimator is formulated in the framework of energy norms, aiming to provide error bounds forthe standard displacement finite element solutions. The formulation is developed initially forlinear elasticity, and subsequently extended to finite elasticity and plasticity. Representativenumerical simulations of these problems are presented, which show the performance of themethod.
2. ENHANCED ASSUMED STRAIN (EAS) FINITE ELEMENT FORMULATION
The EAS finite element formulation [2, 13, 14, 15] is based on the discrete variational equationsobtained from the Hu-Washizu functional [17]. The existence of a function of strain energydensity is assumed, which may be expressed as a function of the linear strain tensor ε ≡ εij
for small strain problems, or the deformation gradient F ≡ FiJ for finite strain problems.The key ingredient of EAS formulations is to enrich the solution as compared to a purely
displacement based formulation, using additional assumed strain modes. For small strains [15]the starting point is an additive decomposition of the strain field into a compatible part andan enhanced part:
ε = ∇su︸︷︷︸comp.
+ ε︸︷︷︸enh.
; εij =12(ui,j + uj,i) + εij , (1)
where ∇su (symmetric component of the displacement gradient) is the compatible part of strainfield, and ε is the enhanced (or incompatible) part. This decomposition allows an enhancementof the numerical solution associated to the incompatible part in discrete models (for the exactcontinuum mechanics solution the field ε would be null). There are no requirements in principleof inter-element continuity for the enhanced field ε, resulting in additional strain modes whichare internal to each element.
For finite deformation [13, 18] EAS may be formulated in terms of displacements anddeformation gradients, parametrised via the following additive decomposition:
F = ∇Xϕ︸ ︷︷ ︸comp.
+ F︸︷︷︸enh.
; FiJ = xi,J + FiJ , (2)
where ϕ : X 7→ x is the deformation mapping, and ∇X is the gradient with respect to theoriginal configuration.
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
2 F. GABALDON AND J.M. GOICOLEA
Full details of EAS formulations may be gathered from the above cited references, and in[6] regarding their implementation for this work.
3. ERROR ESTIMATION BASED ON ENERGY NORMS
3.1. Variational structure of the boundary value problem
We shall consider a boundary value solid mechanics problem, for which the strong formulationis expressed in terms of the unknown displacement field u via a differential operator A(u) inΩ ∪ ∂Ω:
A(u) = 0 ≡
div σ + b = 0 in Ω;u = g in ∂gΩ;σ · n = t in ∂tΩ.
(3)
We shall further assume a variational structure of the problem, which makes (3) equivalent tominimizing the functional Π(u) =
∫Ω
W dΩ−∫Ω
b ·u dΩ−∫∂tΩ
t ·udΓ, where W = W (ε(u),x)is the function of strain energy density, b the volume forces and t the applied surface tractions.
The Dirichlet form a(u)[δu, δu] associated to the strain energy is defined as the secondvariation of the functional. Defining the tangent elastic moduli D = ∂2W/∂ε∂ε, it is expressedas
a(u)[δu, δu] =∫
Ω
∇xδu · (D ·∇xδu)dΩ =
∫
Ω
Dijklδui,jδuk,ldΩ (4)
where sums on repeated indices are understood, and δu ∈ V are admissible variations withinthe space of functions with finite energy norm, satisfying homogeneous kinematic boundaryconditions in ∂gΩ.
For the particular case of linear elasticity D does not depend on u, and application of (4)to the displacement field u itself yields twice the strain energy, a[u, u] =
∫Ω
ε · (D · ε) dΩ =2
∫Ω
W dΩ. In a general case, the energy norm for a given field v is defined as
‖v‖Edef= a(u)[v, v]
12 . (5)
The form a(u)[δu, δu] is regular if 1) the coefficients D ≡ Dijkl are bounded in
Ω, and 2) there exists C ∈ R+ such that a(u)[δu, δu]12 > C‖δu‖1, where ‖δu‖1 def=[∑1
α=0
∫Ω(Dαδu)2 dΩ
] 12
(1st. Sobolev norm). Under these circumstances the functional Π(u)is convex and has a unique relative minimum. Hence, the solution u of the boundary valueproblem verifies:
Π(u) = infv∈V
Π(v). (6)
The variational equation associated to the functional Π(u) is:
G(u)[δu] = 0 ∀δu ∈ V. (7)
where G(u)[δu] is the weak form derived from the boundary value problem (3). Following weshall restrict ourselves to the case of linear elasticity,
G(u)[δu] = −∫
Ω
σ ·∇x(δu) dΩ +∫
Ω
b · δu dΩ +∫
∂tΩ
t · δu dΓ (8)
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 3
Let us now consider the following two expressions: firstly, the restriction of (8) to variationsδuh ∈ Vh,
G(u)[δuh] = 0 ∀δuh ∈ Vh; (9)
and second, the particularisation of (8) to uh ∈ Vh,
G(uh)[δuh] = 0 ∀δuh ∈ Vh. (10)
Subtracting these two equations, taking into account the linearity of G(u)[δu], the followingresult is obtained, expressed in terms of the Dirichlet form:
a(u)[u− uh, δuh] = 0 ∀δuh ∈ Vh (11)
Equation (11) indicates that the finite element solution minimises the value of the energynorm ‖u − uh‖E . This property is referred to as the optimal approximation property of thefinite element method:
‖u− uh‖E = infvh∈Vh
‖u− vh‖E . (12)
3.2. Local error estimation
In general, the finite element solution is obtained in the discretised domain Ωh, which isconstructed via the discretisation of the body Ω ⊂ R3 in nel elements Ωe, e = 1, . . . nel.
Let the finite energy function ue be the exact displacement field in Ωe. The finite dimensioninterpolant polynomial of the exact solution ue is:
ueh(x) def=
nnen∑a=1
uaNa(x) ∈ Pk(Ωe) (13)
being nnen the number of nodes of element e, and Pk(Ωe) the set of polynomials over Ωe withdegree lower or equal than k.
The local error function in element Ωe is defined in each point as the difference betweenthe exact displacement field and the approximate displacement field computed via the finiteelement method:
Ee(x) = ue(x)− ueh(x) (14)
The objective of the local error estimator is to obtain an upper bound of the local errorfunction (14), which may be expressed as:
‖ue − ueh‖E ≤ C(he)α|ue| (15)
where C > 0 is constant, he the diameter of Ωe, |ue| a seminorm of ue and α the rate ofconvergence. The inequality (15) is verified if the optimal approximation property (12) and theregularity conditions for a(u)[·, ·] hold [5].
3.3. Global error estimation
Since ‖uhe‖E is bounded, if the shape functions in (13) are conforming, then the global
interpolant function defined as uh =∑nel
e=1 ueh is also a finite energy function. In such situation,
the upper bound of the global error function expressed in (15) can be obtained as the sum of
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
4 F. GABALDON AND J.M. GOICOLEA
the element contributions. Performing this sum, if the regularity conditions hold, the inequality(15) results in:
‖u− uh‖E ≤ C
nel∑e=1
(he)k+1
ρe|u|k+1 (16)
where |u|k+1 is the (k + 1)st. Sobolev seminorm, |u|k+1 =[∫
Ω(Dk+1u)2 dΩ
] 12 , and ρe is the
diameter of the largest sphere in Ωe.
4. ERROR ESTIMATOR PROPOSED
From a practical point of view, equation (16) is not convenient because the error is expressedin terms of the a priori unknown exact solution ue. Further it is not possible to substitute thisfield by its approximate solution ue
h, as it is a polynomial of degree k and the seminorm usedis of order k + 1 (Dk+1ue
h = 0).Our objective is to build an error estimator for the approximate solution, calculated with
(compatible) displacement elements. For this the finite element solution uenh obtained withthe EAS elements described in section 2 will be used. The starting point is the triangularinequality for the energy norm as follows:
‖u− uh‖E ≤ ‖u− uenh‖E + ‖uenh − uh‖E (17)
Assuming both finite element solutions are convergent, we can express
‖u− uenh‖E = O(hm) (18)‖uenh − uh‖E = O(hp) (19)
A key assumption we shall now make is that, at least in the asymptotic regime (h → 0), therates of convergence in the above expressions are as follows:
m > p (20)
In these conditions, as the mesh is refined, the first term on the right hand side of equation(17) becomes negligible as compared to the second one. Therefore it is possible to establishthat:
‖u− uh‖E ≤ C‖uenh − uh‖E , C ∈ R+ (21)
Hypotheses (18, 19, 20) may be interpreted plainly in the following terms: it is assumedthat both the enhanced solution uenh and the displacement solution uh converge to the exactsolution, in such manner that
1. Their difference, measured by the energy norm ‖uenh − uh‖E , decreases with meshrefinement;
2. The solution obtained with enhanced elements is a better approximation to the exactsolution than the solution with standard displacement elements to that obtained withenhanced elements (m > p).
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 5
The local error estimator proposed is the energy norm of the difference between the enhancedand the displacement solutions:
(Ee)2 = ‖ueenh − ue
h‖E (22)
As discussed in subsection 3.3, the global error may then be obtained as the sum of the localerrors:
E2 =nel∑e=1
(Ee)2 (23)
In summary, in order to compute the local error estimator (22), the discretisation errorassociated to the standard elements is quantified via the energy norm associated to theincompatible modes computed with enhanced elements.
Each component in the sum (23) is local. Therefore the proposed estimator has the importantadvantage that it is computed element by element, without recourse to global smoothingtechniques nor sub-domain solutions.
5. ERROR ESTIMATION IN FINITE ELASTICITY PROBLEMS
The error estimator proposed above may be generalised for finite elasticity problems withhyperelastic models. Here the unknown field shall be the deformation mapping ϕ : Ω 3 X 7→x ∈ Ωt, where Ω is the reference configuration and Ωt is the deformed configuration at timet. The basis of this formulation is similar to that described in section 3, replacing here thedisplacement field u by the deformation mapping ϕ, and the linear strain tensor ε by thedeformation gradient F = ∇Xϕ.
With respect to the approximation methodology the following result is obtained in a similarway to equation (11):
G(ϕ)[δϕh]−G(ϕh)[δϕh] = 0 ∀δϕh ∈ Vh (24)
In this case it’s not possible to group the two terms in (24) leading to a expression similar to11), because in finite elasticity the weak form G(ϕ)[δϕ] is nonlinear in ϕ. However, consideringthe asymptotic regime (h → 0), the finite element solution ϕh will approach the exact solution;as (ϕ−ϕh) → 0, equation (24) may be linearised, resulting in:
G(ϕ)[δϕh]−G(ϕh)[δϕh] ≈ a(ϕ)[ϕ−ϕh, δϕh] = 0 ∀δϕh ∈ Vh, h → 0 (25)
This condition establishes the optimal approximation property of the finite element method forfinite elasticity, in the asymptotic regime:
‖ϕ−ϕh‖E = infvh∈Vh
‖ϕ− vh‖E h → 0 (26)
For finite deformations, the equivalent assumptions to (18, 19) are expressed now in termsof the deformation mapping ϕ:
‖ϕ−ϕenh‖E = O(hm) (27)‖ϕenh −ϕh‖E = O(hp) (28)
m > p (29)
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
6 F. GABALDON AND J.M. GOICOLEA
Then, if (27, 28, 29) are verified, the expression of the local error estimator proposed in (22)results in:
(Ee)2 = ‖ϕeenh −ϕe
h‖E (30)
For numerical implementation, the value of (30) is computed in the reference configuration.The expression of the energy norm, for a given field value η, is [6]:
(‖η‖E)2 = a(ϕ)[η, η] =∫
Ω
∇xη · (D ·∇xη) dΩ (31)
where D is the tangent tensor of constitutive moduli:
D =∂2W (X, F )
∂F ∂F(32)
Straightforward calculations in (30) taking into account (2) provide the closed form of theerror estimator [6]:
(Ee)2 =12
∫
Ωe
F · (D · F ) dΩ (33)
being F the enhanced part of the deformation gradient (2). Note that expression (33) isconfiguration dependent, as moduli D change. Computing the global error via (33) extendedover the complete domain Ω, it can be expressed as the sum of the local errors:
E2 =nel∑e=1
(Ee)2 (34)
6. ERROR ESTIMATION IN VON MISES PLASTICITY
The methodology of error estimation discussed in section 4 and generalised to finite elasticityproblems in section 5, is now further extended to elastic-plastic small strain problems, withVon Mises yield criterion. The methodology described above assumes a variational structureof the boundary value problem. In plasticity, this variational structure may be obtained at anincremental level via the variational integration of the plasticity equations [10]. This variationalintegration postulates the existence of an incremental energy function per unit volume Wt+∆t,for solving step t + ∆t from the results in step t, such that
σt+∆t =∂Wt+∆t
∂εet+∆t
(35)
where εet+∆t are the elastic strains at time t + ∆t.
For concreteness, in what follows we assume small strain J2 plasticity with isotropichardening. The functional dependence of Wt+∆t is on elastic strain and effective plastic strainξ =
∫ t
02/3 (εp · εp)1/2 dt. The expression of the incremental potential function is:
Wt+∆t(εet+∆t, ξt+∆t, ε
et , ξt) = min
ξt+∆t
(Ψt+∆t(εe
t+∆t, ξt+∆t)−Ψt(εet , ξt)
)(36)
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 7
where Ψ(εe, ξ) is the free energy function. The requirement for a minimum in the right-handside of (36) is equivalent to the condition:
∂Ψt+∆t(εet+∆t, ξt+∆t)
∂ξt+∆t= 0 (37)
Assuming that the elastic response is independent of the phenomena associated tounrecoverable distortions of the crystalline lattice, the free energy function may be expressedvia the additive decomposition in an elastic part and a plastic part. Besides, if the additivedecomposition of the infinitesimal strain tensor is assumed:
Ψ(εe, ξ) = Ψe(εe) + Ψp(ξ); ε = εe + εp, (38)
the incremental potential Wt+∆t can be written as:
Wt+∆t = minξt+∆t
(Ψe
t+∆t(εt+∆t − εpt+∆t) + Ψp
t+∆t(ξt+∆t)−Ψet (εt − εp
t )−Ψpt (ξt)
)(39)
The optimisation condition (37) applied to (39), leads to the following expression [6]:
(J2,t+∆t
)2 =23
∂Ψp
∂ξt+∆t(40)
where J2 is the second invariant of the deviatoric part of the Cauchy stress tensor. In thissituation the boundary value problem is described via an incremental differential operator,At+∆t(u) = 0. The corresponding Dirichlet form is
a(ut+∆t)[δu, δu] =∫
Ω
∇s(δu) ·(
∂2Wt+∆t
∂εt+∆t∂εt+∆t·∇s(δu)
)dΩ (41)
If (41) verifies the regularity conditions, the error estimation methodology described inprevious sections is applicable. Then, the local error estimator for this kind of problems is:
(Ee∆t)
2 = ‖ueenht+∆t
− ueht+∆t
‖E (42)
The error bound proposed in (42) is an incremental bound of absolute error in a giventime-step ∆t. In order to evaluate the discretisation error along the complete load path, it isnecessary to determine the integral of Ee
∆t over time:
(Eet+∆t)
2 =∫ t+∆t
0
(Ee∆t)
2dt (43)
The above expression indicates that accumulation of the error is performed by summing thesquares. This causes one of the limitations of this error estimator as a global measure forthe complete load path: it is computed as an increasing monotonic function, as the integrandis always positive. It will not reflect correctly situations in which errors may tend to reducelocally along a complex load path or bifurcations.
Using the incremental function Wt+∆t, the error estimator is interpreted as the contributionof the incompatible modes of the free energy function:
(Ee∆t)
2 =∫
Ωe
Wt+∆t
(εe
t+∆t − εet+∆t(u), ξt+∆t − ξt+∆t(u), εe
t − εet (u), ξt − ξt(u)
)dΩ (44)
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
8 F. GABALDON AND J.M. GOICOLEA
The energy density in (44) can be decomposed additively with the contributions of theelastic and plastic part of the free energy, resulting in:
(Ee∆t)
2 =∫
Ωe
W et+∆t
(εe
t+∆t − εet+∆t(u), εe
t − εet (u)
)dΩ+
∫
Ωe
W pt+∆t (ξt+∆t − ξt+∆t(u), ξt − ξt(u)) dΩ (45)
The global discretisation error is obtained extending the integral in (44) to the completedomain Ω. Then, the global error is computed via the summation of the local errors:
E2∆t =
nel∑
i=1
(Ei∆t)
2 (46)
7. NUMERICAL SIMULATIONS
7.1. 3-D linear elasticity. Scordelis-Lo roof
The classical Scordelis-Lo test for shell elements [12], is solved here with 3D solid elements inorder to show the behaviour of the error estimator using an isotropic linear elasticity model. Onan 80 cylindrical sector roof a volumetric load is acting in Z direction. The roof is supported ontwo rigid diaphragms. Figure 1 shows the problem definition taking into account the symmetryconditions. The numerical values adopted for Lame elastic moduli are: λ = 0, µ = 2.16 · 108,and the dimensions of the roof are L = 50, R = 25, t = 0.25 and b = 360.
x
y
z
L/2
40
freeC
A
B
D
symm
symm
R
diaphragm
Figure 1. Scordelis-Lo roof. Definition of the problem.
The results obtained with enhanced assumed strain elements are very adequate for errorestimation because they approach closely those obtained with shell elements, as shown intable I for several meshes with only one element across the thickness. The results for shellelements referenced in this table have been obtained with the four node general purpose thickshell element S4 from a commercial program element library [1]. The vertical displacement ofpoint B computed with 16 × 16 × 1 enhanced solid elements is vB = 0.3031, very similar tothe result vref = 0.3024 reported in the literature [12].
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 9
Table I. Results obtained with 3D enhanced elements and thick shell elements for Scordelis-Lo test
No of elements vB (solid elements) vB (shell elements)
4× 4× 1 0.1656 0.31328× 8× 1 0.2896 0.303116× 16× 1 0.3031 0.301632× 32× 1 0.3038 0.301664× 64× 1 0.3042 0.3018
For error estimation the exact solution adopted here is obtained with a 24961 D.O.F. mesh,solved with enhanced assumed strain elements. The error estimation is analysed refining overthe surface of the roof with only one element along the thickness, as well as refining throughthe thickness. The meshes considered for the first case are those referenced in table I. Themeshes used for refinement along the thickness are 16 × 16 × 1, 16 × 16 × 2, 16 × 16 × 3,16× 16× 4, 16× 16× 5 and 16× 16× 10.
Figure 2 shows the energy norm versus the number of degrees of freedom, for refinementover surface. The refinement across the thickness does not show a significant influence on theenergy norm [7]. Figures 3 and 4 show the evolution of the global error estimated in terms of
100000
1e+06
1e+07
1e+08
1e+09
1e+10
100 1000 10000 100000
Ene
rgy
N
||ϕref||E||ϕenh||E||ϕh||E
Figure 2. Variation of energy norm with no. of D.O.F. for refinement over the surface
the number of degrees of freedom. The theoretical error is evaluated as the difference betweenthe reference energy norm and the energy norm computed with standard elements. Finally,figure 5 shows the error contours obtained with two of the considered meshes.
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
10 F. GABALDON AND J.M. GOICOLEA
1
10
100
1000
10000
100000
100 1000 10000 100000
Ene
rgy
N
Theoretical errorEstimated error
Figure 3. Evolution of global error with respect to no. of D.O.F. for refinement over the surface
1
10
100
1000
10000
100000
1000 10000
Ene
rgy
N
Theoretical errorEstimated error
Figure 4. Evolution of global error with no. of D.O.F. for refinement along thickness
8.26E+00
1.37E+01
1.92E+01
2.46E+01
3.01E+01
3.55E+01
2.81E+00
4.10E+01
L O C A L E R R O R
Current ViewMin = 2.81E+00X = 2.28E+01Y = 1.06E+01Z = 0.00E+00Max = 4.10E+01X = 2.51E+01Y = 0.00E+00Z = 2.50E+01
Time = 0.00E+00
8.26E+00
1.37E+01
1.92E+01
2.46E+01
3.01E+01
3.55E+01
2.81E+00
4.10E+01
L O C A L E R R O R
Current ViewMin = 2.81E+00X = 2.28E+01Y = 1.06E+01Z = 0.00E+00Max = 4.10E+01X = 2.51E+01Y = 0.00E+00Z = 2.50E+01
Time = 0.00E+00
1.97E+00
3.48E+00
4.98E+00
6.49E+00
7.99E+00
9.50E+00
4.65E-01
1.10E+01
L O C A L E R R O R
Current ViewMin = 4.65E-01X = 2.32E+01Y = 9.61E+00Z = 0.00E+00Max = 1.10E+01X = 2.49E+01Y = 0.00E+00Z = 2.50E+01
Time = 0.00E+00
1.97E+00
3.48E+00
4.98E+00
6.49E+00
7.99E+00
9.50E+00
4.65E-01
1.10E+01
L O C A L E R R O R
Current ViewMin = 4.65E-01X = 2.32E+01Y = 9.61E+00Z = 0.00E+00Max = 1.10E+01X = 2.49E+01Y = 0.00E+00Z = 2.50E+01
Time = 0.00E+00
Figure 5. Scordelis roof. Contours of iso-error computed for the 8× 8× 1 and 16× 16× 1 meshes
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 11
7.2. 2-D Finite elasticity. Double notched specimen
This test has been proposed for error estimation by Radovitzky and Ortiz [11]. A doublenotched specimen in plain strain is considered in mode I crack opening. The analysis is carriedout with displacement control. The hyperelastic constitutive model has the following energyfunction:
W (C) =12λ(log J)2 − µ log(J) +
12µ(trace(C)− 3) (47)
with C = F T F (right Cauchy tensor), J = det(F ) and (λ, µ) the Lame elastic moduli. Thevalues adopted for these constants are: λ = 164.21 and µ = 80.194. Six meshes have beenconsidered with 16, 64, 256, 1024, 4096 and 6400 elements with uniform remeshing. The lastone is considered in order to obtain a reference value for the energy norm ‖ϕref‖E .
Figure 6 shows the contours of local error for the meshes with 16 and 64 elements. Thesecontours are drawn over the deformed mesh (without magnification) in order to show the finitestrains reached (lips of the notch were initially in touch). As expected, the discretisation erroris concentrated near the tip of the crack.
3.55E-02
6.27E-02
8.98E-02
1.17E-01
1.44E-01
1.71E-01
8.40E-03
1.98E-01
L O C A L E R R O R
Current ViewMin = 8.40E-03X = 0.00E+00Y = 1.20E+00
Max = 1.98E-01X = 1.46E+00Y = 1.82E-01
Time = 1.00E+00
3.55E-02
6.27E-02
8.98E-02
1.17E-01
1.44E-01
1.71E-01
8.40E-03
1.98E-01
L O C A L E R R O R
Current ViewMin = 8.40E-03X = 0.00E+00Y = 1.20E+00
Max = 1.98E-01X = 1.46E+00Y = 1.82E-01
Time = 1.00E+00
1.88E-02
3.66E-02
5.43E-02
7.20E-02
8.97E-02
1.07E-01
1.14E-03
1.25E-01
L O C A L E R R O R
Current ViewMin = 1.14E-03X = 0.00E+00Y = 1.20E+00
Max = 1.25E-01X = 1.21E+00Y = 1.33E-01
Time = 1.00E+00
1.88E-02
3.66E-02
5.43E-02
7.20E-02
8.97E-02
1.07E-01
1.14E-03
1.25E-01
L O C A L E R R O R
Current ViewMin = 1.14E-03X = 0.00E+00Y = 1.20E+00
Max = 1.25E-01X = 1.21E+00Y = 1.33E-01
Time = 1.00E+00
Figure 6. Double notched specimen. Contours of iso-error for 16 and 64 elements meshes
Figures 7 and 8 show the energy norm and global error versus the number of degrees offreedom, respectively. From these figures it is concluded that the error estimator is convergentand it agrees with the theoretical error.
7.3. Plasticity. Square plate with circular hole
This final example simulates the tensile test of a square plate with a circular hole under planestrain conditions. It has been used as benchmark test for error estimation in [16]. A summaryof the results obtained has been published in [4]. There are two singular points in the diameternormal to the direction of the imposed displacements. The layout and boundary conditions areshown in figure 9, where l = 50 and r = 10, and u = 0.32 is the final value of the prescribeddisplacement. The constitutive behaviour is modelled with an elastic-plastic Von Mises model
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
12 F. GABALDON AND J.M. GOICOLEA
5.25
5.5
5.75
6
6.25
6.5
1 10 100 1000 10000 100000
Ene
rgy
N
||ϕref||E||ϕenh||E||ϕh||E
Figure 7. Double notched specimen. Variation of energy norm versus no. of D.O.F.
0.0001
0.001
0.01
0.1
1
10
100
1 10 100 1000 10000 100000
Ene
rgy
N
Theoretical errorEstimated error
Rate of convergence 1/2
Figure 8. Double notched specimen. Variation of energy error versus D.O.F.
with the following saturation law for hardening:
σY = σ0 + (σ∞ − σ0)(1− e−βξ) + Hξ (48)
The numerical values adopted are: σ0 = 450, σ∞ = 750, β = 16.93, H = 129. The elasticbehaviour is defined with elastic moduli E = 206900 and ν = 0.29. For the error estimationfour meshes with 37, 407, 1583 and 3119 degrees of freedom are considered.
Figure 10 shows the curve of energy norm computed with enhanced elements and the globalerror estimated at the end of the computation, with respect to the number of degrees offreedom. The values of the error estimator obtained predict a low order of convergence: the1/8 slope plotted in double logarithmic scale is well adjusted to the rate of convergence obtainedin the computations.
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 13
¶¶7
ÁÀ
¿
?
6
? ? ? ? ? ? ? ? ? ??
6 6 6 6 6 6 6 6 6 66
¾ -
r
u
u
l
l
Figure 9. Square plate with circular hole. Geometry and boundary conditions
4
8
16
10 100 1000 10000
Err
or
N
||u||enh · 103
Estimated errorRate of convergence 1/8
Figure 10. Square plate with circular hole. Global error estimated
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
14 F. GABALDON AND J.M. GOICOLEA
Figure 11 shows the local error contours at the end of the process. These values are obtainedadding along the load path the incremental values obtained via (44). The maximum values ofcontours are reached in the singular point, and as the mesh is refined the contours are localisedin a 45 band.
9.92E-01
1.83E+00
2.67E+00
3.51E+00
4.35E+00
5.19E+00
1.53E-01
6.03E+00
L O C A L E R R O R
Current ViewMin = 1.53E-01X = 1.00E+02Y = 0.00E+00
Max = 6.03E+00X = 1.00E+01Y = 0.00E+00
Time = 1.00E+00
9.92E-01
1.83E+00
2.67E+00
3.51E+00
4.35E+00
5.19E+00
1.53E-01
6.03E+00
L O C A L E R R O R
Current ViewMin = 1.53E-01X = 1.00E+02Y = 0.00E+00
Max = 6.03E+00X = 1.00E+01Y = 0.00E+00
Time = 1.00E+00
2.89E-01
5.70E-01
8.51E-01
1.13E+00
1.41E+00
1.69E+00
8.76E-03
1.97E+00
L O C A L E R R O R
Current ViewMin = 8.76E-03X = 3.45E+01Y = 0.00E+00
Max = 1.97E+00X = 9.80E+00Y = 1.99E+00
Time = 1.00E+00
2.89E-01
5.70E-01
8.51E-01
1.13E+00
1.41E+00
1.69E+00
8.76E-03
1.97E+00
L O C A L E R R O R
Current ViewMin = 8.76E-03X = 3.45E+01Y = 0.00E+00
Max = 1.97E+00X = 9.80E+00Y = 1.99E+00
Time = 1.00E+00
16 elements 192 elements
1.29E-01
2.56E-01
3.83E-01
5.10E-01
6.38E-01
7.65E-01
1.96E-03
8.92E-01
L O C A L E R R O R
Current ViewMin = 1.96E-03X = 0.00E+00Y = 1.70E+01
Max = 8.92E-01X = 1.00E+01Y = 0.00E+00
Time = 1.00E+00
1.29E-01
2.56E-01
3.83E-01
5.10E-01
6.38E-01
7.65E-01
1.96E-03
8.92E-01
L O C A L E R R O R
Current ViewMin = 1.96E-03X = 0.00E+00Y = 1.70E+01
Max = 8.92E-01X = 1.00E+01Y = 0.00E+00
Time = 1.00E+00
1.25E-01
2.49E-01
3.74E-01
4.98E-01
6.23E-01
7.47E-01
2.04E-04
8.72E-01
L O C A L E R R O R
Current ViewMin = 2.04E-04X = 8.06E+01Y = 3.02E+01
Max = 8.72E-01X = 1.00E+01Y = 0.00E+00
Time = 1.00E+00
1.25E-01
2.49E-01
3.74E-01
4.98E-01
6.23E-01
7.47E-01
2.04E-04
8.72E-01
L O C A L E R R O R
Current ViewMin = 2.04E-04X = 8.06E+01Y = 3.02E+01
Max = 8.72E-01X = 1.00E+01Y = 0.00E+00
Time = 1.00E+00
768 elements 1536 elements
Figure 11. Square plate with circular hole. Contours of local error
Finally the evolution of the local error near the singular point is studied. The error isdecomposed additively into an elastic and a plastic part, in accordance to (45). The twocomponents are shown in figure 12 and 13, respectively. The conclusions obtained from thesecurves are:
1. The elastic and plastic parts of the local error decrease with mesh refinement, althoughthe plastic part is activated earlier in the refined meshes.
2. Both components of the error increase along the load process3. The order of magnitude of both components is similar.
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
ERROR ESTIMATION BASED ON ENHANCED STRAIN ELEMENTS 15
0
5
10
15
20
25
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Err
or
Displacement
16 Elements192 Elements768 Elements
1536 Elements
Figure 12. Square plate with a hole. Evolution of the elastic part of local error near the singular point.
0
5
10
15
20
25
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Err
or
Displacement
16 Elements192 Elements768 Elements
1536 Elements
Figure 13. Square plate with a hole. Evolution of the plastic part of local error near the singular point.
8. CONCLUDING REMARKS
A methodology of error estimation for non linear problems has been described. The errorestimator proposed is based on the energy norm contribution of incompatible modes and inconsequence, the estimated error is zero for the the patch test strain modes. It has been appliedto linear elasticity, non-linear finite elasticity and von Mises elastic-plastic problems.
The estimator evaluates the error for a standard displacement Galerkin finite element model.For this, it applies an auxiliary enriched solution with assumed strain modes.
The formulation results in an algorithm which is evaluated locally, element by element,
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls
16 F. GABALDON AND J.M. GOICOLEA
without need to global equations nor smoothing techniques. This makes it particularly simpleand inexpensive.
The limitations of the proposed approach arise from several factors. Firstly, a key assumptionis made with regard to convergence rates (20), for which we do not provide a rigorous proof.Secondly, the error estimate for nonlinear problems is computed as a monotonic increasingfunction, which precludes valid error estimates for certain load paths. In particular, no attemptis done to discuss bifurcations or instabilities.
In spite of the above limitations, the representative numerical examples presented show agood performance in practical terms, which is encouraging.
REFERENCES
1. ABAQUS. Theory Manual. Hibbitt, Karlsson and Sorensen, Inc, 1998.2. Armero F, Glaser S. On the formulation of enhanced strain finite elements in finite deformations.
Engineering Computations 1997; 14:759–791.3. BabuskaI, Rheinboldt W. A posteriori error estimates for the finite element method. International Journal
for Numerical Methods in Engineering 1978; 12:1597–1613.4. Barthold F, Schmidt M, Stein E. Error indicators and mesh refinements for finite-element-computations of
elastoplastic deformations. Computational Mechanics 1998; 22(3):225–238.5. Ciarlet P. The finite element method for elliptic problems. North Holland, 1978.6. Gabaldon F. Metodos de elementos finitos mixtos con deformaciones supuestas en elastoplasticidad. Ph.D.
thesis, E.T.S. Ingenieros de Caminos, Canales y Puertos. Universidad Politecnica de Madrid. 1999.7. Gabaldon F and Goicolea, J. Error estimation based on non-linear enhanced assumed strain elements. In
Proocedings of ECCOMAS 2000. 11–14 September. Barcelona, Spain. 2000.8. Johnson C, Hansbo P. Adaptive finite element methods in computational mechanics. Computer Methods
in Applied Mechanics and Engineering 1992; 101:143–181.9. Ortiz M, Quigley J. Adaptive mesh refinement in strain localisation problems. Computer Methods in Applied
Mechanics and Engineering 1991; 90:781–804.10. Ortiz M, Stainier L. The variational formulation of viscoplastic constitutive updates. Computer Methods
in Applied Mechanics and Engineering 1999; 171(3–4):419–44411. Radovitzky R, Ortiz M. Error estimation and adaptive meshing in strongly nonlinear dynamic problems.
Computer Methods in Applied Mechanics and Engineering 1998; 172(1–4):203–240.12. Scordelis A, Lo K. Computer analysis of cylindrical shells. Journal of American Concrete Institute 1969;
61:539–561.13. Simo J, Armero F. Geometrically nonlinear enhanced strain mixed methods and the method of incompatible
modes. International Journal for Numerical Methods in Engineering 1992; 33:1413–1449.14. Simo J, Armero F, Taylor, R. Improved versions of assumed enhanced strain tri-linear elements for 3d finite
deformation problems. Computer Methods in Applied Mechanics and Engineering 1993; 110:359–386.15. Simo J, Rifai, S. A class of mixed assumed methods and the method of incompatible modes. International
Journal for Numerical Methods in Engineering 1990; 29:1595–1638.16. Stein E, Schmidt M, Barthold F. Theorie und algorithmen adaptiver fe-methoden fur elastoplastische
deformationen. In Adaptive finite element methoden in der angewandten mechanik . 1997.17. Washizu K. Variational Methods in Elasticity & Plasticity (3st edn). Pergamon Press, 1982.18. Korelc J, Wriggers P. Consistent gradient formulation for a stable enhanced strain method for large
deformations. Engineering Computations 1996; 13:103–123.
Copyright c© 2000 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2000; 00:0–0Prepared using nmeauth.cls